Professional Documents
Culture Documents
Sciencedirect: Enhancing Digital Human Motion Planning of Assembly Tasks Through Dynamics and Optimal Control
Sciencedirect: Enhancing Digital Human Motion Planning of Assembly Tasks Through Dynamics and Optimal Control
com
ScienceDirect
Procedia CIRP 44 (2016) 20 – 25
Abstract
Better operator ergonomics in assembly plants reduce work related injuries, improve quality, productivity and reduce cost. In this paper we
investigate the importance of modeling dynamics when planning for manual assembly operations. We propose modeling the dynamical human
motion planning problem using the Discrete Mechanics and Optimal Control (DMOC) method, which makes it possible to optimize with respect to
very general objectives. First, two industrial cases are simulated using a quasi-static inverse kinematics solver, demonstrating problems where
this approach is sufficient. Then, the DMOC-method is used to solve for optimal trajectories of a lifting operation with dynamics. The resulting
trajectories are compared to a steady state solution along the same path, indicating the importance of using dynamics.
©c 2016
2016 The Authors.
Authors. Published
Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license
(http://creativecommons.org/licenses/by-nc-nd/4.0/).
Peer-review under responsibility of the organizing committee of the 6th CIRP Conference on Assembly Technologies and Systems (CATS).
Peer-review under responsibility of the organizing committee of the 6th CIRP Conference on Assembly Technologies and Systems (CATS)
Keywords: Assembly; Digital human modeling; Ergonomy; Dynamics; Optimal control
1. Introduction hands and feet matches certain target frames. For the quasi-static
inverse kinematics this leads to an underdetermined system of
Although the degree of automation is increasing in manu- equations since the number of joints exceeds the end-effectors
facturing industries, many assembly operations are performed constraints. Due to this redundancy there exist a set of solutions,
manually. To avoid injuries and to reach sustainable production allowing us to consider ergonomics aspects, collision avoidance,
of high quality, comfortable environments for the operators are and maximizing comfort when choosing one solution.
vital, see [1] and [2]. Poor station layouts, poor product de- The dynamic motion planning problem is stated as an optimal
signs or badly chosen assembly sequences are common sources control problem, which we discretize using discrete mechanics.
leading to unfavorable poses and motions. To keep costs low, This results in an optimization problem, which can be solved
preventive actions should be taken early in a project, raising the using standard nonlinear programming solvers. Furthermore,
need for feasibility and ergonomics studies in virtual environ- this general problem formulation makes it fairly easy to include
ments long before physical prototypes are available. very general constraints and objectives.
Today, in the automotive industries, such studies are con- In this paper we show, using a couple of case studies, where
ducted to some extent. The full potential, however, is far from the quasi-static solver is sufficient, and where the DMOC solver
reached due to limited software support in terms of capability could improve the solution. The paper extends the work pre-
for realistic pose prediction, motion generation and collision sented in [4] and [5], and is a part of Cromm (Creation of Muscle
avoidance. As a consequence, ergonomics studies are time con- Manikins) project [6].
suming and are mostly done for static poses, not for full assembly
motions. Furthermore, these ergonomic studies, even though
performed by a small group of highly specialized simulation 2. Background
engineers, show low reproducibility within the group [3].
To describe operations and facilitate motion generation, it is 2.1. Manikin Model
common to equip the manikin with coordinate frames attached to
end-effectors like hands and feet. The inverse kinematic problem In this section we present the manikin model and the inverse
is to find joint values such that the position and orientation of kinematic problems, both quasi-static and with dynamics.
2212-8271 © 2016 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license
(http://creativecommons.org/licenses/by-nc-nd/4.0/).
Peer-review under responsibility of the organizing committee of the 6th CIRP Conference on Assembly Technologies and Systems (CATS)
doi:10.1016/j.procir.2016.02.125
Staffan Björkenstam et al. / Procedia CIRP 44 (2016) 20 – 25 21
2.4. Discrete mechanics and optimal control This principle is equivalent to the discrete Euler-Lagrange
equations:
2.4.1. The constrained discrete Euler-Lagrange equations
D2 Ld (q(k−1) , q(k) , h) + D1 Ld (q(k) , q(k+1) , h)
Consider the mechanical system specified by a configuration
+ F+d (q(k−1) , q(k) , u(k−1) , h) + F−d (q(k) , q(k+1) , u(k) , h)
manifold Q ⊆ Rnq and Lagrangian L : T Q → R, where T Q is
the tangent bundle of the configuration manifold. Furthermore, + ΦT (q(k) )λ(k)
d = 0, (8a)
suppose the motion of the system is constrained by the equation φ(q (k+1)
) = 0, (8b)
φ(q) = 0 ∈ Rm to lie in the constraint manifold C = φ−1 (0) ⊂ Q.
Let U ∈ Rnu be the set of admissible controls and F : T Q × U → where D1 Ld and D2 Ld are the slot derivatives with respect to
T ∗ Q the external force acting on the system, where T ∗ Q is the the first and second argument. These equations define the varia-
cotangent bundle of the configuration manifold. tional integrator by implicitly mapping (q(k−1) , q(k) , u(k−1) , u(k) )
Introducing the multiplier λ(t) ∈ Rm the Lagrange- to (q(k+1) , λ(k)
d ). Please refer to [11] for a thorough introduction to
d’Alembert principle states that trajectories of the system satisfy discrete mechanics and [12,13] for more on discrete mechanics
and optimal control of multibody systems.
t2 A reasonable trade-off between accuracy and performance, is
δ L(q(t), q̇(t)) + φT (q(t))λ(t)dt to use the the midpoint rule to approximate the relevant integrals.
t1
t2 The discrete Lagrangian then becomes
+ F(q(t), q̇(t), u(t)) · δqdt = 0, (5) q + q q − q
0 1 1 0
t1 Ld (q0 , q1 , h) = hL , . (9)
2 h
where variations are taken with respect to q, fixed at the end-
points, and with respect to λ. Thus
Integration by parts and the fundamental lemma of calculus h ∂L q0 + q1 q1 − q0
D1 Ld (q0 , q1 , h) = ,
of variations give the following differential algebraic equations, 2 ∂q 2 h
known as the constrained Euler Lagrange equations of motion: ∂L q0 + q1 q1 − q0
− ,
∂L d ∂L ∂ q̇ 2 h
(q(t), q̇(t)) − (q(t), q̇(t))
∂q dt ∂ q̇ and
+ F(q(t), q̇(t), u(t)) + ΦT (q(t))λ(t) = 0,
h ∂L q0 + q1 q1 − q0
(6a)
φ(q(t)) = 0, (6b) D2 Ld (q0 , q1 , h) = ,
2 ∂q 2 h
where Φ denotes the Jacobian of the constraint function. ∂L q0 + q1 q1 − q0
+ , .
The key idea of variational integrators is to directly approx- ∂ q̇ 2 h
imate the variational principle (5) rather than the equations of Furthermore, it is then natural to use the following discrete
motion (6). forces:
We now discretize q(t) in [t1 , t2 ] using a fixed time step h =
(t2 − t1 )/N so that q(k) is an approximation of q(t1 + kh) for F+d (q0 , q1 , u0 , h) = F−d (q0 , q1 , u0 , h) =
k = 0, . . . , N. Furthermore, we discretize the control such that h q0 + q1 q1 − q0
u(k) is an approximation of u(t1 + (k + 12 )h) for k = 0, . . . , N − 1. = F , , u0 . (10)
2 2 h
We are now ready to replace the continuous state space, T Q,
with the discrete state space, Q × Q, and construct a discrete This discretization scheme results in a second order accurate
Lagrangian Ld : Q × Q × R → R such that integrator.
t1 +(k+1)h
2.5. Optimal control problem
Ld (q(k) , q(k+1) , h) ≈ L(q(t), q̇(t))dt.
t1 +kh
We consider the following optimal control problem: Mini-
Introducing left and right discrete forces, F+d and F−d , and mize
discrete multipliers, λ(k)
d for k = 0, . . . , N, a discrete variational tf
principle corresponding to (5) can be formulated as J = χ(q(t f ), q̇(t f )) + L(q(t), q̇(t), u(t))dt (11a)
t0
N−1
δ (Ld (q , q
(k) (k+1)
, h) subject to
k=0
∂L d ∂L
1
+ φT (q(k) )λ(k)
1 T (k+1) (k+1) (q(t), q̇(t)) − (q(t), q̇(t))
2 d + 2 φ (q )λd ) ∂q dt ∂ q̇
N−1 + F(q(t), q̇(t), u(t)) + ΦT (q(t))λ(t) = 0, (11b)
+ (F−d (q(k) , q(k+1) , u(k) , h) · δq(k) φ(q(t)) = 0, (11c)
k=0
g(q(t), q̇(t), u(t)) ≥ 0, (11d)
+ F+d (q(k) , q(k+1) , u(k) , h) · δq(k+1) ) = 0 (7)
ψ0 (q(t0 ), q̇(t0 )) = 0, (11e)
for all variations δλ(k)
d and δq
(k)
with δq(0) = δq(N) = 0. ψ f (q(t f ), q̇(t f )) = 0 (11f)
Staffan Björkenstam et al. / Procedia CIRP 44 (2016) 20 – 25 23
for t ∈ [t0 , t f ].
Thus, we want to minimize a performance index (11a), con-
sisting of the terminal cost, χ, and the integral of the control
Lagrangian, L, along the trajectory, while satisfying the dynam-
ics (11b)-(11c), path constraints (11d), and boundary conditions
(11e)-(11f).
It is well known that the discrete mechanics formulation of
the equations of motion show excellent conservation of quanti-
ties, such as momenta and energy, conserved by the continuous (a) Start (b) Enter
system. This will enable us to take larger time steps and still
get physically meaningful results [14]. There is, however, yet
another computational advantage when used in optimal control.
Namely, since there are no explicit references to velocities in the
discrete equations of motion, the resulting optimization problem
can be formulated using fewer variables, compared to standard
discretizations of trajectories on T Q.
Approximating the objective using the midpoint rule and en-
forcing the path constraints at the midpoints we get the following
discrete optimal control problem: Minimize (c) Finishing (d) End
f (x) (13a) The second case is to place washers inside the trunk of a car,
this case is also provided by Volvo Cars. The case can be divided
such that the constraints into two steps: first place the washers, and then to mount the
bolts. Since both steps require the same reachability and force
cl ≤ c(x) ≤ cu (13b) we choose to simulate only the washer placing. The manikin
uses the left hand as support on the trunk floor to extend the
and simple bounds reach, and the hand is free to rotate on that surface. The case
is tried with 8 different manikins to cover the anthropometric
xl ≤ x ≤ xu (13c) variables length and weight, and also both sexes.
24 Staffan Björkenstam et al. / Procedia CIRP 44 (2016) 20 – 25
(a) (b)
χ(q, q̇) = 0
L(q, q̇, u) = uT u
ψ0 (q, q̇) = q̇
ψ f (q, q̇) = q̇
Fig. 3: Top: The reachability for the shortest manikin is insufficient
Bottom: The end is in reach for the manikin
where the control signal, u, is chosen to be the normalized
actuator torque. The problem is then solved for both a 10 kg and
Staffan Björkenstam et al. / Procedia CIRP 44 (2016) 20 – 25 25
1 Acknowledgements